@agentworkforce/harness-kit

Composable primitives for spawning a persona's harness (claude, codex, opencode) with its MCP servers, env vars, and permissions wired up correctly.

This is the layer that @agentworkforce/cli sits on top of. If you're building your own orchestrator on top of @agentworkforce/workload-router and want the same behaviors the CLI provides — env-ref resolution, MCP isolation, permission flag translation — depend on this package rather than reimplementing them.

The router (@agentworkforce/workload-router) models what a persona is. This kit models how to launch it on a given harness. Both are harness-agnostic on their own; the per-harness knowledge lives here.

Install

pnpm add @agentworkforce/harness-kit @agentworkforce/workload-router

What's in the box

buildInteractiveSpec(input) — translate a persona to an interactive argv

Takes the fields off a PersonaSelection (harness, model, systemPrompt, harnessSettings, mcpServers, permissions) and returns {bin, args, initialPrompt, warnings}. Pure — no I/O, no stderr writes. Warnings are returned so your caller routes them wherever makes sense.

import { resolvePersona } from '@agentworkforce/workload-router';
import {
  buildInteractiveSpec,
  resolveMcpServersLenient,
  resolveStringMapLenient,
  formatDropWarnings
} from '@agentworkforce/harness-kit';
import { spawn } from 'node:child_process';

const selection = resolvePersona('persona-authoring');

// Resolve env + MCP refs against the current process environment. Missing
// refs don't throw — they come back on `.dropped` so you can warn the user.
const envResolution = resolveStringMapLenient(selection.env, process.env, 'env');
const mcpResolution = resolveMcpServersLenient(selection.mcpServers, process.env);

const warnings = formatDropWarnings(
  envResolution.dropped,
  mcpResolution.dropped,
  mcpResolution.droppedServers
);
for (const w of warnings) console.warn(w);

// Build the exec spec.
const spec = buildInteractiveSpec({
  harness: selection.runtime.harness,
  personaId: selection.personaId,
  model: selection.runtime.model,
  systemPrompt: selection.runtime.systemPrompt,
  harnessSettings: selection.runtime.harnessSettings,
  mcpServers: mcpResolution.servers,
  permissions: selection.permissions
});
for (const w of spec.warnings) console.warn(w);

// Spawn the harness.
const args = spec.initialPrompt ? [...spec.args, spec.initialPrompt] : [...spec.args];
spawn(spec.bin, args, {
  stdio: 'inherit',
  env: { ...process.env, ...(envResolution.value ?? {}) }
});

useRunnablePersona(intent) — run a persona non-interactively

For orchestrators that need a programmatic sendMessage() surface, the kit also exposes a thin runner around the same router + harness translation path. It resolves the persona, launches the selected harness in non-interactive mode, captures stdout/stderr, reports progress chunks, supports cancellation and timeouts, and returns a stable execution result.

useRunnablePersona follows the router's internal built-in resolver. For optional pack/local personas, resolve a PersonaSelection through your source cascade and call useRunnableSelection(selection).

import { useRunnablePersona } from '@agentworkforce/harness-kit';

const persona = useRunnablePersona('persona-authoring');
const run = persona.sendMessage('Draft a persona for workflow artifact writing.', {
  workingDirectory: process.cwd(),
  name: 'persona-author',
  timeoutSeconds: persona.selection.runtime.harnessSettings.timeoutSeconds,
  inputs: {
    TARGET_DIR: '.agentworkforce/workforce/personas',
    CREATE_MODE: 'local',
    TASK_DESCRIPTION: 'Write a workflow artifact as structured JSON.'
  },
  onProgress: (chunk) => process.stderr.write(chunk.text)
});

const result = await run;
if (result.status !== 'completed') {
  throw new Error(result.stderr || `persona run failed: ${result.status}`);
}
console.log(result.output);

If the selected persona declares inputs, sendMessage(..., { inputs }) resolves those values before spawn, substitutes $NAME / ${NAME} in the system prompt, and injects the resolved values into the child process env. Resolution uses explicit inputs, then process.env[spec.env ?? NAME], then default, and throws when a required input is still unset.

The runner maps harnesses to their non-interactive command shapes: claude --print, codex exec, and opencode run. It writes generated config files such as opencode.json only for the duration of the child process and restores or removes them afterward. Skill installation is opt-in with installSkills: true; callers that need stronger filesystem isolation should keep using a mount/sandbox layer around the runner.

Claude harness guarantees

When harness === 'claude', buildInteractiveSpec always emits both:

• --mcp-config '{"mcpServers": …}' — even if empty
• --strict-mcp-config — forces Claude Code to ignore user/project MCP sources

This means a persona session only sees MCP servers the persona itself declares. Your ~/.claude.json and any project .claude/ MCP config are invisible inside the session. That's the whole point of persona isolation; if you want a personal MCP in the session, declare it on the persona.

Codex / opencode

Current state: these harnesses don't expose runtime MCP injection or permission controls on their CLIs. buildInteractiveSpec carries the system prompt as the initial positional [PROMPT] argument (since neither has a --system-prompt flag) and returns a warning string if the persona declares mcpServers or permissions. The caller decides whether to print, fail, or continue.

Env reference resolution

The kit supports two forms of env references inside persona JSON:

Unbraced $VAR mid-string is kept as a literal — this prevents a stray $ in a JSON value from accidentally being treated as a reference, and it keeps missing-var errors pointed at a specific field name.

Two resolution policies

Pick the one that matches your error-handling preference:

The CLI uses the lenient path; it drops missing env entries and unset MCP headers with a warning, and only aborts if a structural field (url, command, any arg) can't be resolved.

Persona input rendering

Persona inputs are distinct from env references. Inputs are prompt-visible runtime values declared on a persona, such as TARGET_DIR, PACKAGE_NAME, or CREATE_MODE. Use them for non-secret launch context, not API keys.

import { renderPersonaInputs, resolvePersonaInputs } from '@agentworkforce/harness-kit';

const { values } = resolvePersonaInputs(
  {
    TARGET_DIR: { env: 'MY_TARGET_DIR', default: './out' },
    CREATE_MODE: { default: 'local' }
  },
  { TARGET_DIR: '/tmp/personas' },
  process.env
);

const systemPrompt = renderPersonaInputs(
  'Write to $TARGET_DIR using ${CREATE_MODE} mode.',
  values
);

resolvePersonaInputs fails hard for missing required inputs. That is intentional: unlike secret env refs, an input is usually structural context the persona needs to follow its contract.

API surface

// Persona inputs
export class MissingPersonaInputError extends Error { input: string; env: string }
export function resolvePersonaInputs(inputs, provided, processEnv): PersonaInputResolution
export function renderPersonaInputs(systemPrompt, values): string
export interface PersonaInputResolution { values: Record<string,string> }
export type PersonaInputValues = Record<string, string | number | boolean | null | undefined>

// Env refs
export class MissingEnvRefError extends Error { ref: string; referencedBy: string }
export function makeEnvRefResolver(env): (value, field) => string
export function makeLenientResolver(env): (value, field) => LenientResult
export function resolveStringMap(map, env, prefix): Record<string,string> | undefined
export function resolveStringMapLenient(map, env, prefix): { value, dropped: DroppedRef[] }
export type LenientResult = { ok: true; value: string } | { ok: false; field: string; ref: string }
export interface DroppedRef { field: string; ref: string }

// MCP
export function resolveMcpServersLenient(servers, env): McpResolution
export function formatDropWarnings(envDrops, mcpDrops, mcpServerDrops): string[]
export interface McpResolution {
  servers: Record<string, McpServerSpec> | undefined;
  dropped: DroppedRef[];
  droppedServers: DroppedMcpServer[];
}
export interface DroppedMcpServer { name: string; refs: string[] }

// Harness
export function buildInteractiveSpec(input: BuildInteractiveSpecInput): InteractiveSpec
export interface BuildInteractiveSpecInput {
  harness: Harness;
  personaId: string;
  model: string;
  systemPrompt: string;
  harnessSettings?: HarnessSettings;
  mcpServers?: Record<string, McpServerSpec>;
  permissions?: PersonaPermissions;
}
export interface InteractiveSpec {
  bin: string;
  args: readonly string[];
  initialPrompt: string | null;
  warnings: string[];
}

// Runnable personas
export function useRunnablePersona(intent, options?): RunnablePersonaContext
export function useRunnableSelection(selection, options?): RunnablePersonaContext
export interface RunnablePersonaContext {
  selection: PersonaSelection;
  install: PersonaInstallContext;
  sendMessage(task, options?): PersonaExecution;
}
export interface PersonaExecutionResult {
  status: 'completed' | 'failed' | 'cancelled' | 'timeout';
  output: string;
  stderr: string;
  exitCode: number | null;
  durationMs: number;
}

Status

Small, stable surface focused on the three things a harness spawner needs: resolve env refs, resolve MCP config, and build argv. The default exports are still pure when you use buildInteractiveSpec directly. The useRunnablePersona convenience is intentionally the small side-effecting layer for consumers that want the same harness knowledge plus a captured non-interactive child process.